Span evals
person concept tool org talk claim — click a node to jump to its page; hover an arrow for the relation
A span eval is a type of evals|evaluation scoped to a single component (or "span") within an LLM pipeline, as opposed to evaluations that require data aggregated across multiple components.
Definition and scope
Dat Ngo of Arize describes span evals as evaluations that "can be run on a single kind of component," where that component is referred to as a "span." (Dat Ngo, 17:00) This makes them the atomic unit of evaluation within an Observability|LLM observability framework — assessing one discrete step or node in a larger pipeline rather than the system as a whole.
Contrast with multi-span evals
Span evals are explicitly contrasted with Multi-span evals, which Ngo describes as necessary when "the eval that I want... actually requires data across many different components." (Dat Ngo, 17:32) The distinction maps onto granularity: span evals are local and self-contained, while multi-span evals aggregate context across the trace to support evaluations whose criteria cannot be satisfied by any single component in isolation.
Role in observability
Span evals are positioned as a building block of LLM observability pipelines. By enabling evaluation at the individual component level, they allow teams to isolate which specific step in a multi-step agent or RAG pipeline is responsible for quality issues, rather than only assessing end-to-end outputs.